Research map / Adversarial robustness and security / Certified and empirical robustness
Intrinsic Explanation of Random Subspace Method for Enhanced Security Applications
Yanting Wang, Jinyuan Jia
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EnsembleSHAP instead of Shapley value | random subspace ensemble (BERT) · IMDb (no attack, delete 10%) | faithfulness (label flip rate): 0.300 -> 0.600 | Improved |
| use EnsembleSHAP instead of Shapley value | random subspace ensemble (BERT) · IMDb (adversarial attack, delete 10%) | faithfulness (label flip rate): 0.845 -> 0.980 | Improved |
| use EnsembleSHAP instead of Shapley value | random subspace ensemble (BERT) · AG-news (backdoor attack, delete 10%) | faithfulness (label flip rate): 0.725 -> 0.735 | No clear effect |
| use EnsembleSHAP instead of LIME | random subspace ensemble (BERT) · SST-2 (no attack, delete 10%) | faithfulness (label flip rate): 0.125 -> 0.365 | Improved |
| use EnsembleSHAP instead of Shapley value | RA-LLM (Vicuna-7B) · DAN jailbreak (delete 10%) | faithfulness (label flip rate): 0.33 -> 0.85 | Improved |
| use EnsembleSHAP instead of LIME | RA-LLM (Vicuna-7B) · GCG jailbreak (delete 10%) | faithfulness (label flip rate): 0.15 -> 0.15 | No clear effect |
| use EnsembleSHAP instead of Shapley value | random subspace ensemble (BERT) · IMDb (backdoor attack, top-5 words) | key-word recall: 0.491 -> 0.892 | Improved |
| use EnsembleSHAP instead of Shapley value | random subspace ensemble (BERT) · SST-2 (adversarial attack, top-5 words) | key-word precision: 0.361 -> 0.378 | No clear effect |
| use EnsembleSHAP instead of LIME | RA-LLM (Vicuna-7B) · AutoDAN jailbreak (top-10 words) | key-word precision: 0.335 -> 0.434 | Improved |
| use EnsembleSHAP instead of Shapley value | RA-LLM (Vicuna-7B) · DAN jailbreak (top-10 words) | key-word recall: 0.289 -> 0.287 | No clear effect |
| increase number of sub-sampled inputs N (ablation) | EnsembleSHAP · text classification | faithfulness and key-word prediction | Improved |
| increase dropping rate to a very large value (0.9) (ablation) | EnsembleSHAP · text classification | faithfulness | Hurt |
Closest papers by meaning
- Adversarial Robustness via Deformable Convolution with Stochasticity (ICML 2025 · similarity 0.83)
- Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness (NeurIPS 2024 · similarity 0.83)
- Rethinking Robustness of Model Attributions (AAAI 2024 · similarity 0.82)
- One Stone, Two Birds: Enhancing Adversarial Defense Through the Lens of Distributional Discrepancy (ICML 2025 · similarity 0.82)
- Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning (CVPR 2024 · similarity 0.82)
- Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
